Answers are grounded in this paper and cite their sources.
Fault identification in renewable-energy-integrated distribution networks is primarily challenged by two coupled issues: learning discriminative representations from dynamically reconfigured, high-order network topologies and maintaining robust classification under limited or imbalanced fault labels, measurement disturbances, and incomplete observations. To address these challenges, this study proposes a dynamic hypergraph diffusion prototype network, termed DHDP-Net. The framework consists of a dynamic hypergraph feature extraction module, a masked temporal diffusion encoding module, and a soft-threshold prototype identification module. The dynamic hypergraph module integrates physical connectivity with data-driven node correlations to capture high-order spatial dependencies among buses, feeders, and power-supply regions under topology variations. The masked temporal diffusion module combines masked reconstruction with a temporal transformer denoiser to learn normal temporal evolution patterns and extract fault-sensitive residuals from reconstruction deviations. Learnable soft thresholding suppresses weak residual responses caused by renewable energy fluctuations and measurement disturbances, while prototype matching constructs stable category representations in a discriminative metric space, improving fault identification when labeled samples are limited or unevenly distributed. Experiments on the IEEE 123-bus system and SMART-DS show that DHDP-Net achieves Macro F1 scores of 98.41 and 97.02%, respectively. Under degraded measurement conditions with a 30% measurement-missing ratio, the proposed model retains Macro F1 scores of 95.82 and 95.48%, outperforming the strongest baseline by 3.51 and 4.62 percentage points, respectively. Ablation, computational-efficiency, and parameter-sensitivity studies further verify the effectiveness of the proposed architecture. These results demonstrate that DHDP-Net provides a topology-aware and robust solution for fault identification under limited-label and degraded-measurement conditions.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: